Pith. sign in

REVIEW 4 major objections 5 minor 60 references

Replacing a U-Net decoder's upsampling stage with a wavelet-domain suppression block preserves fine detail and cuts computation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 22:19 UTC pith:4I54IPHI

load-bearing objection A plausible decoder block with one solid empirical result, but the cost claim is contradicted by the paper's own limitation section. the 4 major comments →

arxiv 2511.11015 v2 pith:4I54IPHI submitted 2025-11-14 cs.CV

SUPER Module for Detail-Sensitive and Cost-Efficient U-Net Variant Decoders

Siheon Joo , Hongjo Kim This is my paper
classification cs.CV
keywords U-Net decoderperfect reconstructiondiscrete wavelet transformselective suppressionhigh-frequency detailthin-crack segmentationimage denoisingdepth estimation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

SUPER is a plug-in decoder block that aims to fix the detail loss common in U-Net-style decoders without forcing the whole network into a wavelet-specific architecture. The block applies a discrete wavelet transform to the skip feature, learns to suppress redundant subband content, and reconstructs via the inverse transform; when the suppression term is zero, the block reconstructs the skip feature exactly. On the paper's tests, the block raises edge accuracy in monocular depth estimation and thin-crack IoU for sub-2px cracks while reducing decoder multiply-accumulate operations. The authors argue this is a structural improvement, not a tuning artifact: the same block, with encoders and losses unchanged, benefits both a high-frequency task (thin cracks) and a low-frequency task (smartphone denoising).

Core claim

SUPER's central claim is that the lossy spatial upsampling step in U-Net decoders can be replaced by a reconstruction-aware wavelet-domain operation. At decoder stage k the block computes x_d^k = W^T(W(x_e^k) - F_d^k(W(x_e^k) (+) x_d^{k+1})), where W is the discrete wavelet transform and F_d^k is a learnable decoder block (CBAM attention in the default implementation). If F_d^k is zero, the stage is exactly W^T W x_e^k = x_e^k, the wavelet perfect-reconstruction identity. The learned nonzero term selectively suppresses redundant components in the wavelet domain, giving each stage the form ΨΦ = I - S with a bounded operator S. The paper reports that installing this block inside two existing U

What carries the argument

The load-bearing object is Eq. (3), the SUPER block recurrence x_d^k = W^T(W(x_e^k) - F_d^k(W(x_e^k) (+) x_d^{k+1})). DWT/IDWT act as the perfect-reconstruction pair; the learnable decoder transformation F_d^k plays the role of a suppression operator S in the identity ΨΦ = I - S, so zero suppression reproduces the input exactly. In the practical block, F_d^k is a small convolutional/attention stack (CBAM in the default implementation) that operates on all four wavelet subbands, and fusion with the deeper stage happens entirely in the low-frequency LL band to avoid interpolation artifacts. The theoretical apparatus is the operator-norm bound ‖S‖ ≤ ε < 1, which yields the multi-stage stability

Load-bearing premise

The whole stability argument assumes the learned suppression operator shrinks redundant frequency content with norm below 1, but nothing in the implementation measures or enforces that bound.

What would settle it

Measure the largest amplification factor (spectral norm) of the trained suppression operator at each decoder stage; if any stage can amplify rather than shrink, the stability bound ‖ΨΦ‖ ≤ (1+ε)^L is violated. As a second check, replace the DWT/IDWT pair with any fixed invertible transform and retrain; if the gains vanish, the mechanism is wavelet-specific, and if not, the benefit comes from the residual structure rather than perfect reconstruction.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If SUPER's central claim holds, any skip-connected U-Net-like decoder can regain high-frequency fidelity without retraining the encoder or changing the task head.
  • Fine-structure metrics—not just averages—should improve: the paper reports the largest gains in exactly the 0–2 px crack-width regime where conventional decoders fail.
  • Decoder computational cost drops because the wavelet transform is MAC-neutral and the block avoids both spatial upsampling and the channel compression that upsampling decoders use.
  • The two components are mutually dependent in practice: removing either CBAM or the suppression path led to non-convergence in all reported runs.
  • Low-frequency tasks are not hurt: on denoising, the block matched or slightly improved PSNR/SSIM while cutting decoder work by two-thirds.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to medical imaging or remote sensing, where the same sub-4px failure regime appears in micro-lesion or fine-line detection; the paper's logic predicts SUPER helps wherever thin structures are the deciding signal.
  • The theory does not depend on the Haar wavelet, so testing non-Haar or learned wavelet bases could separate the perfect-reconstruction property from the specific basis; the paper only validated Haar.
  • The authors note they never estimated the spectral norm of the trained suppression operator; measuring it stage-by-stage would test whether the stability bound is actually satisfied in practice.
  • Combining SUPER with task-specific losses (e.g., topology-aware crack loss) reportedly destabilized training at default hyperparameters, suggesting the block changes decoder activation distributions enough that loss retuning is part of the integration recipe.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SUPER, a plug-in decoder block that replaces the upsampling/refinement stages of U-Net-like decoders with a wavelet-domain operation. At each stage, the skip feature is decomposed by DWT, fused with the deeper decoder output at reduced resolution, passed through a learnable block (with CBAM), and subtracted before IDWT. The authors argue this preserves fine high-frequency structure via perfect reconstruction while suppressing redundant components, and they claim an improved detail–cost trade-off. Experiments are reported on CrackVision12K thin-crack segmentation with FACS-Net and SIDD denoising with CascadedGaze-Net; the abstract also reports iDisc depth results and decoder MAC reductions, though these are absent from the main text.

Significance. A plug-and-play decoder block that improves detail fidelity without increasing cost would be a useful contribution, and the wavelet-domain formulation is a reasonable design direction. The paper makes explicit falsifiable claims, including specific edge-AUC and MAC-reduction numbers. However, the cost half of the central claim is unsupported and contradicted by the paper's own limitation section; the theoretical 'SUPER' property is a definitional identity rather than a derived guarantee; and the implementation does not operationalize the selective-suppression subspace of the theory. The one in-body positive result (Table 1) compares against a BCE-retrained baseline without error bars, and the SIDD gain is within noise. As presented, the central detail–cost trade-off is not established.

major comments (4)
  1. [Sec. 6.4 vs. Abstract; Sec. 4.3] The cost-efficiency claim is both unmeasured and internally contradicted. The title and abstract promise decoder MAC reductions (iDisc 12.69G→4.36G; SIDD 66.6%), but §5 contains no MAC table and no iDisc experiment. §4.3's 'Computational Analysis' only asserts MACs(H,W,C)≈MACs(H/2,W/2,4C), which is a statement about the transform, not about the CBAM/DC/NAF blocks actually used. Worse, §6.4 states 'While the overall MACs remain unchanged, parameter growth slightly increases representational capacity,' which directly contradicts the abstract. The cost-efficiency half of the paper's central claim therefore has no measured support and is internally inconsistent.
  2. [Sec. 3, Eq. (5)] The SUP ER property is tautological. The paper defines Ψ_kΦ_k = I − S_k and then presents this as a 'relaxation' of perfect reconstruction; but S_k is not independently characterized—any residual decoder trivially satisfies the identity. The contraction assumption ‖S‖≤ε<1 is 'imposed' by fiat (Sec. 3) rather than enforced by spectral normalization or projection, and ε is never estimated. Consequently the multi-stage bound ‖ΨΦ‖≤(1+ε)^L is a restatement of the assumption, not a stability guarantee. The theoretical foundation does not distinguish SUPER from any other residual decoder.
  3. [Sec. 4.1, Algorithm 1] Selective suppression is not implemented as theorized. The theory assumes an orthogonal split x=x_T+x_S and a suppression operator acting only on the redundant subspace 𝒮=𝒯⊥. Algorithm 1 computes out = IDWT(bands − chunk(res)), where res = CBAM(F_d^k(W(x_e^k) ⊕ x_d^{k+1})); there is no projection onto 𝒯 or 𝒮, and the subtracted term acts on all wavelet subbands. Thus the claimed 'conditional PR' and 'selective suppression' properties are not properties of the implemented block. Moreover, Eq. (4)'s exact-reconstruction claim requires W^TW=I, which holds only for orthonormal wavelets with specific boundary handling; the implementation uses Haar but also channel chunking/fusion, so the 'reduces to x_e^k' statement is not automatic.
  4. [Sec. 5.1, Table 1; Sec. 5.2, Table 2] Empirical support for the detail claim is weak in the body. Table 1 reports SUPER-FACS (BCE) IoU 0.451 vs FACS-Net (BCE) 0.356 for 0–2 px cracks, but this is not the published FACS-Net, which uses CT-Loss; the authors state they could not stably train SUPER with CT-Loss under default hyperparameters. No error bars or repeated-seed statistics are given. The ablation statement that removing either CBAM or SS led to 'gradient vanishing in all ten runs' does not isolate component contributions; it indicates optimization failure, not that each component is necessary for the stated mechanism. Table 2's PSNR gain over CascadedGaze (40.416 vs 40.390) is 0.026 dB, likely within run-to-run variability. The abstract's iDisc numbers appear nowhere in the body.
minor comments (5)
  1. [Figure 1] Caption contains typo 'Deocder'. Also the main text and abstract disagree on whether SUPER reduces MACs or keeps them comparable; please harmonize.
  2. [Appendix A] Appendix A states it 'will be released in a future update'; no implementation details, code, or hyperparameters are provided, so the reproducibility claim in Sec. 5 is not met.
  3. [References] Reference [42] is cited for the MAC-neutrality of wavelets but is a ChIP-seq peak-calling tool; this citation appears to be a placeholder. Refs [24] and [28] are duplicate entries for the same FACS-Net paper.
  4. [Sec. 3, Eq. (5)] Notation: 𝒮 is used both for the redundant subspace and for the suppression operator; Eq. (5) writes |S_k|≤ε where an operator norm is needed.
  5. [Algorithm 1] The operation 'chunk(res)' is not defined; clarify how the 4C-channel residual is split across subbands and how fusion with x_d^{k+1} (sum vs concatenation) is chosen.

Circularity Check

1 steps flagged

Theoretical core is definitional: Eq. (5) restates S = I − ΨΦ and the contraction bound is imposed, not enforced; empirical benchmarks remain independent.

specific steps
  1. self definitional [Sec. 3 and Sec. 4.1, Eq. (5)]
    "To explicitly model selective reconstruction, we introduce the Selectively Suppressed Perfect Reconstruction (SUPER) operator: ΨΦ = I − S, ‖S‖ ≤ ε, where S is a learnable suppression operator acting on redundant components... Formally, each SUPER block satisfies the suppressed reconstruction property Ψ_k Φ_k = I − S_k, |S_k| ≤ ε, where Φ_k and Ψ_k denote the encoder and decoder frame operators at stage k, and S_k represents the learnable selective suppression in the wavelet domain."

    S is never specified independently of the residual I − ΨΦ; the claimed 'suppressed reconstruction property' is therefore a restatement of the definition of S, not a derived theorem. Any residual decoder stage can be written as ΨΦ = I − S, so the relaxation of perfect reconstruction has no independent content unless S is shown to be the implemented subtraction W^T F_d W restricted to the nuisance subspace. The paper additionally imposes ‖S‖_2 ≤ ε < 1 by fiat and never enforces it in Algorithm 1, where the subtracted term is CBAM(F_d(...)) applied to all wavelet subbands. The multi-stage stability bound (1+ε)^L thus only exponentiates the imposed assumption, making the theoretical 'guarantee' an input rather than a derived prediction.

full rationale

The main derivation-chain problem is localized to the theoretical foundation: Eq. (5) is true by construction because S_k is effectively defined as I − Ψ_kΦ_k, and the contraction bound is assumed rather than enforced by the architecture. This makes the paper's 'relaxed reconstruction framework' largely a relabeling of a residual wavelet decoder. However, the empirical claims are not circular: CrackVision12K and SIDD results are compared against independent baselines (including NAFNet, Restormer, etc.), and the reported IoU/PSNR gains are measured rather than derived from the tautological equation. The self-citations to the authors' own FACS-Net are used as a comparison baseline, not as a load-bearing justification of SUPER's mechanism, so they do not raise the score further. Separately, the cost claim is internally inconsistent: the abstract promises large decoder-MAC reductions, while Sec. 6.4 states 'While the overall MACs remain unchanged, parameter growth slightly increases representational capacity.' That is a correctness/reporting contradiction, not a circular derivation, so it is noted here but not scored as a circular step.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

The central claim leans on: (1) an unestimated, unenforced bound epsilon that carries the entire stability theory; (2) an orthogonal informative/redundant decomposition that the implementation never projects onto; (3) an asserted contraction property of a learned network with no architectural mechanism; plus the standard DWT perfect-reconstruction fact, which is legitimate background. The ledger shows the theoretical content is mostly notation and postulates; the empirical content is the measurements, which are only partially present in the text.

free parameters (3)
  • epsilon (suppression bound)
    Sec. 3 imposes ||S||2 <= epsilon < 1 by fiat; epsilon is never estimated from weights, never enforced by spectral normalization or Lipschitz control, and carries the entire multi-stage stability bound ||Psi Phi|| <= (1+epsilon)^L.
  • Haar wavelet basis (fixed)
    Acknowledged in Sec. 6.4 as 'tested only on a Haar basis'; the central results depend on this hand-chosen fixed basis rather than a learned or adaptive one.
  • Subband chunk/channel allocation in Algorithm 1
    Algorithm 1 step 4 'chunk(res)' and the channel-projection rules for fusion are unspecified; the claimed MAC counts and reconstruction behavior depend on these choices.
axioms (5)
  • standard math Orthonormal DWT satisfies perfect reconstruction W^T W = I with appropriate boundary handling
    Invoked in Sec. 4.1 Eq. (4) and Sec. 3 to establish the 'exact reconstruction when suppression is zero' fallback; this is a legitimate mathematical fact.
  • domain assumption The signal decomposes orthogonally into informative vs redundant subspaces: x = x_T + x_S, <x_T, x_S> = 0
    Sec. 3 postulates this split and the projection P_T, but no data-dependent construction of T is given for any task, and Algorithm 1 never computes P_T.
  • ad hoc to paper The learned suppression operator S is a contraction on the redundant subspace: ||S|| <= epsilon < 1
    Sec. 3 asserts boundedness 'to ensure that nuisance components cannot be amplified'; the implementation's S = I - W^T F_d^k W acts on all subbands, and no architectural mechanism enforces contraction.
  • domain assumption Deep networks exhibit spectral bias favoring low frequencies
    Sec. 1 and Sec. 4.3 rely on this cited prior result ([10], [11]) as motivation for the frequency-aware design; it is background literature, not derived here.
  • domain assumption LL-band fusion is alias-free because the Nyquist frequency of the lower-resolution feature matches the LL bandwidth
    Sec. 4.3 asserts physical consistency of LL-only fusion without derivation or experimental verification of alias-freeness.
invented entities (2)
  • SUPER suppression operator S (and per-stage S_k) no independent evidence
    purpose: Encodes the deviation from perfect reconstruction; basis of the stability analysis in Sec. 3 and Eq. (5).
    It is a notational device: defining S := I - Psi Phi makes 'Psi Phi = I - S' tautological, and the contraction bound is neither measured nor enforced anywhere in the implementation.
  • Conditional PR projection P_T and the informative/redundant subspace split (T, S) no independent evidence
    purpose: Defines task-relevant vs redundant content to justify selective suppression.
    No operational handle: the paper provides no construction of T for crack segmentation, denoising, or depth, and Algorithm 1 never projects onto T; the entity does no work outside the prose.

pith-pipeline@v1.3.0-alltime-deepseek · 14437 in / 27002 out tokens · 239417 ms · 2026-08-03T22:19:01.103122+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of SUPER Module for Detail-Sensitive and Cost-Efficient U-Net Variant Decoders." pith.science (2026). https://pith.science/paper/4I54IPHI

@misc{pith2026251111015,
  author       = {Pith},
  title        = {Pith review of: SUPER Module for Detail-Sensitive and Cost-Efficient U-Net Variant Decoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4I54IPHI}},
  note         = {Machine review of arXiv:2511.11015}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Skip-connected U-Net variants are widely used for dense inverse problems, yet their decoders commonly recover resolution through spatial upscaling, which can blur or distort fine structures. Wavelet transforms provide an explicit perfect-reconstruction path, but prior wavelet networks often couple this property to wavelet-specific architectures or fixed output representations. We propose Selectively Suppressed Perfect Reconstruction (SUPER), a wavelet-domain decoder formulation that replaces unstructured spatial refinement with bounded frequency suppression. In its idealized equal-channel form, SUPER reduces to exact reconstruction when the suppression term is zero; in practical U-Net decoders, it provides a projected-subspace reconstruction fallback with learnable task-specific suppression. The resulting SUPER Module is a structurally plug-in decoder block: it replaces the upsampling/refinement stage of an existing U-Net-like decoder, while leaving the encoder and task head unchanged. We evaluate SUPER on monocular depth estimation, thin-crack segmentation, and smartphone image denoising. On iDisc depth estimation, SUPER improves the host model's edge AUC from 26.59% to 34.45% while reducing decoder MACs from 12.69G to 4.36G. On FACS-Net crack segmentation, SUPER improves average IoU and the extreme 0-2px crack regime. On SIDD denoising, where high-frequency enhancement is less directly rewarded, SUPER preserves PSNR/SSIM while reducing decoder MACs by 66.6%. These results support SUPER as a practical wavelet-domain suppression module for improving the observed detail-cost trade-off of U-Net-like decoders.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 17 canonical work pages · 4 internal anchors

  1. [1]

    A Survey of Deep Learning Video Super- Resolution,

    A. A. Baniya, T.-K. Lee, P. Eklund, and S. Aryal, “A Survey of Deep Learning Video Super- Resolution,” IEEE Trans. Emerg. Top. Comput. Intell., vol. 8, no. 4, pp. 2655–2676, Aug. 2024, doi: 10.1109/TETCI.2024.3398015

    arXiv 2024

  2. [2]

    Deep learning for efficient high- resolution image processing: A systematic review,

    A. Dede et al., “Deep learning for efficient high- resolution image processing: A systematic review,” Intell. Syst. Appl., vol. 26, p. 200505, June 2025, doi: 10.1016/j.iswa.2025.200505

    arXiv 2025

  3. [3]

    Deep Convolutional Framelets: A General Deep Learning Framework for Inverse Problems,

    J. C. Ye, Y. Han, and E. Cha, “Deep Convolutional Framelets: A General Deep Learning Framework for Inverse Problems,” SIAM J. Imaging Sci., vol. 11, no. 2, pp. 991–1048, Jan. 2018, doi: 10.1137/17M1141771

    work page doi:10.1137/17m1141771 2018

  4. [4]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” May 18, 2015, arXiv: arXiv:1505.04597. doi: 10.48550/arXiv.1505.04597

    Pith/arXiv arXiv doi:10.48550/arxiv.1505.04597 2015

  5. [5]

    A Comprehensive Review of U-Net and Its Variants: Advances and Applications in Medical Image Segmentation,

    W. Jiangtao, N. I. R. Ruhaiyem, and F. Panpan, “A Comprehensive Review of U-Net and Its Variants: Advances and Applications in Medical Image Segmentation,” IET Image Process., vol. 19, no. 1, p. e70019, 2025, doi: 10.1049/ipr2.70019

    doi:10.1049/ipr2.70019 2025

  6. [6]

    Multi-level Wavelet-CNN for Image Restoration

    P. Liu, H. Zhang, K. Zhang, L. Lin, and W. Zuo, “Multi-level Wavelet-CNN for Image Restoration,” May 22, 2018, arXiv: arXiv:1805.07071. doi: 10.48550/arXiv.1805.07071

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1805.07071 2018

  7. [7]

    Augmenting Perceptual Super-Resolution via Image Quality Predictors

    F. Zhang, S. B. Rangrej, T. Aumentado-Armstrong, A. Fazly, and A. Levinshtein, “Augmenting Perceptual Super-Resolution via Image Quality Predictors”

  8. [8]

    Next-Gen Medical Imaging: U-Net Evolution and the Rise of Transformers,

    C. Zhang, X. Deng, and S. H. Ling, “Next-Gen Medical Imaging: U-Net Evolution and the Rise of Transformers,” Sensors, vol. 24, no. 14, p. 4668, July 2024, doi: 10.3390/s24144668

    work page doi:10.3390/s24144668 2024

  9. [9]

    Catch Missing Details: Image Reconstruction with Frequency Augmented Variational Autoencoder,

    X. Lin, Y. Li, J. Hsiao, C. Ho, and Y. Kong, “Catch Missing Details: Image Reconstruction with Frequency Augmented Variational Autoencoder,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada: IEEE, June 2023, pp. 1736–1745. doi: 10.1109/CVPR52729.2023.00173

    arXiv 2023

  10. [10]

    On the Spectral Bias of Neural Networks,

    N. Rahaman et al., “On the Spectral Bias of Neural Networks,” in Proceedings of the 36th International Conference on Machine Learning, PMLR, May 2019, pp. 5301–5310. Accessed: Mar. 04, 2025. [Online]. Available: https://proceedings.mlr.press/v97/rahaman19a.html

    2019

  11. [11]

    Towards Understanding the Spectral Bias of Deep Learning,

    Y. Cao, Z. Fang, Y. Wu, D.-X. Zhou, and Q. Gu, “Towards Understanding the Spectral Bias of Deep Learning,” Oct. 05, 2020, arXiv: arXiv:1912.01198. doi: 10.48550/arXiv.1912.01198

    Pith/arXiv arXiv doi:10.48550/arxiv.1912.01198 2020

  12. [12]

    Wavelet U-Net and the Chromatic Adaptation Transform for Single Image Dehazing,

    H.-H. Yang and Y. Fu, “Wavelet U-Net and the Chromatic Adaptation Transform for Single Image Dehazing,” in 2019 IEEE International Conference on Image Processing (ICIP), Sept. 2019, pp. 2736–

    2019

  13. [13]

    Wavelet Domain Style Transfer for an Effective Perception-Distortion Tradeoff in Single Image Super-Resolution,

    X. Deng, R. Yang, M. Xu, and P. L. Dragotti, “Wavelet Domain Style Transfer for an Effective Perception-Distortion Tradeoff in Single Image Super-Resolution,” in 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea (South): IEEE, Oct. 2019, pp. 3076–3085. doi: 10.1109/ICCV.2019.00317

    arXiv 2019

  14. [14]

    Addressing Spectral Bias of Deep Neural Networks by Multi-Grade Deep Learning

    R. Fang and Y. Xu, “Addressing Spectral Bias of Deep Neural Networks by Multi-Grade Deep Learning”

  15. [15]

    WRANet: wavelet integrated residual attention U-Net network for medical image segmentation,

    Y. Zhao, S. Wang, Y. Zhang, S. Qiao, and M. Zhang, “WRANet: wavelet integrated residual attention U-Net network for medical image segmentation,” Complex Intell. Syst., vol. 9, no. 6, pp. 6971–6983, Dec. 2023, doi: 10.1007/s40747- 023-01119-y

    work page doi:10.1007/s40747- 2023

  16. [16]

    Wavelet U-Net++ for accurate lung nodule segmentation in CT scans: Improving early detection and diagnosis of lung cancer,

    S. Akila Agnes, A. Arun Solomon, and K. Karthick, “Wavelet U-Net++ for accurate lung nodule segmentation in CT scans: Improving early detection and diagnosis of lung cancer,” Biomed. Signal Process. Control, vol. 87, p. 105509, Jan. 2024, doi: 10.1016/j.bspc.2023.105509

    arXiv 2024

  17. [17]

    Task-Driven Wavelets Using Constrained Empirical Risk Minimization,

    E. Marcus, R. Sheombarsing, J.-J. Sonke, and J. Teuwen, “Task-Driven Wavelets Using Constrained Empirical Risk Minimization,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA: IEEE, June 2024, pp. 24098–24107. doi: 10.1109/CVPR52733.2024.02275

    arXiv 2024

  18. [18]

    Focal Frequency Loss for Image Reconstruction and Synthesis,

    L. Jiang, B. Dai, W. Wu, and C. C. Loy, “Focal Frequency Loss for Image Reconstruction and Synthesis,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, QC, Canada: IEEE, Oct. 2021, pp. 13899–13909. doi: 10.1109/ICCV48922.2021.01366

    arXiv 2021

  19. [19]

    Optimizing transformer-based network via advanced decoder design for medical image segmentation,

    W. Yang et al., “Optimizing transformer-based network via advanced decoder design for medical image segmentation,” Biomed. Phys. Eng. Express, vol. 11, no. 2, p. 025024, Mar. 2025, doi: 10.1088/2057-1976/adaec7

    work page doi:10.1088/2057-1976/adaec7 2025

  20. [20]

    The Role of Deep Learning in Medical Image Inpainting: A Systematic Review,

    J. C. Santos, H. Tomás Pereira Alexandre, M. Seoane Santos, and P. Henriques Abreu, “The Role of Deep Learning in Medical Image Inpainting: A Systematic Review,” ACM Trans. Comput. Healthc., vol. 6, no. 3, pp. 1–24, July 2025, doi: 10.1145/3712710

    doi:10.1145/3712710 2025

  21. [21]

    3D medical image segmentation using the serial–parallel convolutional neural network and transformer based on cross‐ window self‐attention,

    B. Yu, Q. Zhou, L. Yuan, H. Liang, P. Shcherbakov, and X. Zhang, “3D medical image segmentation using the serial–parallel convolutional neural network and transformer based on cross‐ window self‐attention,” CAAI Trans. Intell. Technol., vol. 10, no. 2, pp. 337–348, Apr. 2025, doi: 10.1049/cit2.12411

    work page doi:10.1049/cit2.12411 2025

  22. [22]

    On understanding and overcoming spectral biases of deep neural network learning methods for solving PDEs,

    Z.-Q. J. Xu, L. Zhang, and W. Cai, “On understanding and overcoming spectral biases of deep neural network learning methods for solving PDEs,” Jan. 17, 2025, arXiv: arXiv:2501.09987. doi: 10.48550/arXiv.2501.09987

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.09987 2025

  23. [23]

    Deep learning in crack detection: A comprehensive scientometric review,

    Y. Wu, S. Li, J. Li, Y. Yu, J. Li, and Y. Li, “Deep learning in crack detection: A comprehensive scientometric review,” J. Infrastruct. Intell. Resil., vol. 4, no. 3, p. 100144, Sept. 2025, doi: 10.1016/j.iintel.2025.100144

    arXiv 2025

  24. [24]

    Frequency-Aware Crack Segmentation Network (FACS-Net) for Thin-Cracks via Topology Preservation

    S. Joo, S. Kim, and H. Kim, “Frequency-Aware Crack Segmentation Network (FACS-Net) for Thin-Cracks via Topology Preservation”

  25. [25]

    Adaptive Feature Medical Segmentation Network: an adaptable deep learning paradigm for high-performance 3D brain lesion segmentation in medical imaging,

    A. Zaman et al., “Adaptive Feature Medical Segmentation Network: an adaptable deep learning paradigm for high-performance 3D brain lesion segmentation in medical imaging,” Front. Neurosci., vol. 18, p. 1363930, Apr. 2024, doi: 10.3389/fnins.2024.1363930

    arXiv 2024

  26. [26]

    Frequency‐aware denoising using a diffusion model for enhanced band‐limited and white noise removal in x‐ray acoustic computed tomography,

    J. Peng, M. Lu, B. Li, J. Wang, W. Hu, and X. Liu, “Frequency‐aware denoising using a diffusion model for enhanced band‐limited and white noise removal in x‐ray acoustic computed tomography,” Med. Phys., vol. 52, no. 5, pp. 3325–3335, May 2025, doi: 10.1002/mp.17681

    work page doi:10.1002/mp.17681 2025

  27. [27]

    Exploring a Frequency- Domain Attention-Guided Cascade U-Net: Towards Spatially Tunable Segmentation of Vasculature,

    N. Mu, Z. Lyu, X. Zhang, R. McBane, A. S. Pandey, and J. Jiang, “Exploring a Frequency- Domain Attention-Guided Cascade U-Net: Towards Spatially Tunable Segmentation of Vasculature,” Comput. Biol. Med., vol. 167, p. 107648, Dec. 2023, doi: 10.1016/j.compbiomed.2023.107648

    arXiv 2023

  28. [28]

    Frequency-Aware Crack Segmentation Network (Facs-Net) for Thin- Cracks Via Topology Preservation,

    S. Joo, S. Kim, and H. Kim, “Frequency-Aware Crack Segmentation Network (Facs-Net) for Thin- Cracks Via Topology Preservation,” Aug. 08, 2025, Social Science Research Network, Rochester, NY: 5384487. doi: 10.2139/ssrn.5384487

    work page doi:10.2139/ssrn.5384487 2025

  29. [29]

    CascadedGaze: Efficiency in Global Context Extraction for Image Restoration,

    A. Ghasemabadi, M. K. Janjua, M. Salameh, C. Zhou, F. Sun, and D. Niu, “CascadedGaze: Efficiency in Global Context Extraction for Image Restoration,” May 07, 2024, arXiv: arXiv:2401.15235. doi: 10.48550/arXiv.2401.15235

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.15235 2024

  30. [30]

    Assessing the Image Quality of Digitally Reconstructed Radiographs from Chest CT,

    O. T. Paalvast, O. Hertgers, M. Sevenster, and H. J. Lamb, “Assessing the Image Quality of Digitally Reconstructed Radiographs from Chest CT,” J. Imaging Inform. Med., Jan. 2025, doi: 10.1007/s10278-025-01406-9

    work page doi:10.1007/s10278-025-01406-9 2025

  31. [31]

    On Removing Interpolation and Resampling Artifacts in Rigid Image Registration,

    I. Aganj, B. T. T. Yeo, M. R. Sabuncu, and B. Fischl, “On Removing Interpolation and Resampling Artifacts in Rigid Image Registration,” IEEE Trans. Image Process., vol. 22, no. 2, pp. 816–827, Feb. 2013, doi: 10.1109/TIP.2012.2224356

    arXiv 2013

  32. [32]

    How Convolutional Neural Networks Deal with Aliasing

    A. H. Ribeiro and T. B. Schön, “How Convolutional Neural Networks Deal with Aliasing,” Feb. 15, 2021, arXiv: arXiv:2102.07757. doi: 10.48550/arXiv.2102.07757

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2102.07757 2021

  33. [33]

    Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection,

    C. Tan et al., “Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection,” Dec. 20, 2023, arXiv: arXiv:2312.10461. doi: 10.48550/arXiv.2312.10461

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.10461 2023

  34. [34]

    An Effective UNet Using Feature Interaction and Fusion for Organ Segmentation in Medical Image

    X. Gou, C. Liao, J. Zhou, F. Ye, and Y. Lin, “FIF- UNet: An Efficient UNet Using Feature Interaction and Fusion for Medical Image Segmentation,” Sept. 09, 2024, arXiv: arXiv:2409.05324. doi: 10.48550/arXiv.2409.05324

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2409.05324 2024

  35. [35]

    LMSC-UNet: A Lightweight U-Net with Modified Skip Connections for Semantic Segmentation:,

    S. Sawant, A. Medgyesy, S. Raghunandan, and T. Götz, “LMSC-UNet: A Lightweight U-Net with Modified Skip Connections for Semantic Segmentation:,” in Proceedings of the 17th International Conference on Agents and Artificial Intelligence, Porto, Portugal: SCITEPRESS - Science and Technology Publications, 2025, pp. 726–734. doi: 10.5220/0013343800003890

    work page doi:10.5220/0013343800003890 2025

  36. [36]

    WST: Wavelet- Based Multi-scale Tuning for Visual Transfer Learning

    J. Zeng, L. Huang, and K. Wang, “WST: Wavelet- Based Multi-scale Tuning for Visual Transfer Learning”

  37. [37]

    Towards Building More Robust Models with Frequency Bias,

    Q. Bu, D. Huang, and H. Cui, “Towards Building More Robust Models with Frequency Bias,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV), Paris, France: IEEE, Oct. 2023, pp. 4379–4388. doi: 10.1109/ICCV51070.2023.00406

    arXiv 2023

  38. [38]

    Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains,

    M. Tancik et al., “Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2020, pp. 7537–7547. Accessed: Oct. 10, 2025. [Online]. Available: https://papers.neurips.cc/paper_files/paper/2020/has h/55053683268957697aa39fba6f231c68- Abstract.html

    arXiv 2020

  39. [39]

    Fourier Neural Operator for Parametric Partial Differential Equations,

    Z. Li et al., “Fourier Neural Operator for Parametric Partial Differential Equations,” May 17, 2021, arXiv: arXiv:2010.08895. doi: 10.48550/arXiv.2010.08895

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.08895 2021

  40. [40]

    Implicit Neural Representations with Periodic Activation Functions,

    V. Sitzmann, J. Martel, A. Bergman, D. Lindell, and G. Wetzstein, “Implicit Neural Representations with Periodic Activation Functions,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2020, pp. 7462–7473. Accessed: Oct. 10, 2025. [Online]. Available: https://proceedings.neurips.cc/paper/2020/hash/53c 04118df112c13a8c34b38343...

    2020

  41. [41]

    CBAM: Convolutional Block Attention Module,

    S. Woo, J. Park, J.-Y. Lee, and I. S. Kweon, “CBAM: Convolutional Block Attention Module,” in Computer Vision – ECCV 2018, vol. 11211, V. Ferrari, M. Hebert, C. Sminchisescu, and Y. Weiss, Eds., in Lecture Notes in Computer Science, vol. 11211. , Cham: Springer International Publishing, 2018, pp. 3–19. doi: 10.1007/978-3-030-01234-2_1

    doi:10.1007/978-3-030-01234-2_1 2018

  42. [42]

    Model-based Analysis of ChIP- Seq (MACS),

    Y. Zhang et al., “Model-based Analysis of ChIP- Seq (MACS),” Genome Biol., vol. 9, no. 9, p. R137, Sept. 2008, doi: 10.1186/gb-2008-9-9-r137

    doi:10.1186/gb-2008-9-9-r137 2008

  43. [43]

    Hybrid-Segmentor: Hybrid approach for automated fine-grained crack segmentation in civil infrastructure,

    J. M. Goo, X. Milidonis, A. Artusi, J. Boehm, and C. Ciliberto, “Hybrid-Segmentor: Hybrid approach for automated fine-grained crack segmentation in civil infrastructure,” Autom. Constr., vol. 170, p. 105960, Feb. 2025, doi: 10.1016/j.autcon.2024.105960

    arXiv 2025

  44. [44]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA: IEEE, June 2016, pp. 770–778. doi: 10.1109/CVPR.2016.90

    doi:10.1109/cvpr.2016.90 2016

  45. [45]

    SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers,” Oct. 28, 2021, arXiv: arXiv:2105.15203. doi: 10.48550/arXiv.2105.15203

    Pith/arXiv arXiv doi:10.48550/arxiv.2105.15203 2021

  46. [46]

    Evaluating Road Crack Segmentation Performance in Participatory Sensing: An Exploration of Alternative Metrics,

    J. Hyeon, M. Jeong, W.-C. Chern, V. K. Asari, and H. Kim, “Evaluating Road Crack Segmentation Performance in Participatory Sensing: An Exploration of Alternative Metrics,” J. Comput. Civ. Eng., vol. 39, no. 5, p. 04025064, Sept. 2025, doi: 10.1061/JCCEE5.CPENG-6011

    work page doi:10.1061/jccee5.cpeng-6011 2025

  47. [47]

    A High- Quality Denoising Dataset for Smartphone Cameras,

    A. Abdelhamed, S. Lin, and M. S. Brown, “A High- Quality Denoising Dataset for Smartphone Cameras,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT: IEEE, June 2018, pp. 1692–1700. doi: 10.1109/CVPR.2018.00182

    arXiv 2018

  48. [48]

    NTIRE 2019 Challenge on Real Image Denoising: Methods and Results,

    A. Abdelhamed et al., “NTIRE 2019 Challenge on Real Image Denoising: Methods and Results,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Long Beach, CA, USA: IEEE, June 2019, pp. 2197–2210. doi: 10.1109/CVPRW.2019.00273

    arXiv 2019

  49. [49]

    Benchmarking Denoising Algorithms with Real Photographs,

    T. Plotz and S. Roth, “Benchmarking Denoising Algorithms with Real Photographs,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI: IEEE, July 2017, pp. 2750–2759. doi: 10.1109/CVPR.2017.294

    work page doi:10.1109/cvpr.2017.294 2017

  50. [50]

    Learning Enriched Features for Real Image Restoration and Enhancement,

    S. W. Zamir et al., “Learning Enriched Features for Real Image Restoration and Enhancement,” in Computer Vision – ECCV 2020, vol. 12370, A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds., in Lecture Notes in Computer Science, vol. 12370. , Cham: Springer International Publishing, 2020, pp. 492–511. doi: 10.1007/978-3-030-58595- 2_30

    work page doi:10.1007/978-3-030-58595- 2020

  51. [51]

    Addressing Spectral Bias of Deep Neural Networks by Multi-Grade Deep Learning

    R. Fang and Y. Xu, “Addressing Spectral Bias of Deep Neural Networks by Multi-Grade Deep Learning,” Oct. 21, 2024, arXiv: arXiv:2410.16105. doi: 10.48550/arXiv.2410.16105

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.16105 2024

  52. [52]

    Simple Baselines for Image Restoration,

    L. Chen, X. Chu, X. Zhang, and J. Sun, “Simple Baselines for Image Restoration,” in Computer Vision – ECCV 2022, vol. 13667, S. Avidan, G. Brostow, M. Cissé, G. M. Farinella, and T. Hassner, Eds., in Lecture Notes in Computer Science, vol. 13667. , Cham: Springer Nature Switzerland, 2022, pp. 17–33. doi: 10.1007/978-3-031-20071-7_2

    doi:10.1007/978-3-031-20071-7_2 2022

  53. [53]

    A lightweight semantic segmentation method for concrete bridge surface diseases based on improved DeeplabV3+,

    Z. Yu, C. Dai, X. Zeng, Y. Lv, and H. Li, “A lightweight semantic segmentation method for concrete bridge surface diseases based on improved DeeplabV3+,” Sci. Rep., vol. 15, no. 1, p. 10348, Mar. 2025, doi: 10.1038/s41598-025-95518-5

    work page doi:10.1038/s41598-025-95518-5 2025

  54. [54]

    A dual encoder crack segmentation network with Haar wavelet-based high–low frequency attention,

    J. Zhang, Z. Zeng, P. K. Sharma, O. Alfarraj, A. Tolba, and J. Wang, “A dual encoder crack segmentation network with Haar wavelet-based high–low frequency attention,” Expert Syst. Appl., vol. 256, p. 124950, Dec. 2024, doi: 10.1016/j.eswa.2024.124950

    arXiv 2024

  55. [56]

    Cross Aggregation Transformer for Image Restoration

    Z. Chen, Y. Zhang, J. Gu, Y. Zhang, L. Kong, and X. Yuan, “Cross Aggregation Transformer for Image Restoration”

  56. [57]

    MAXIM: Multi-Axis MLP for Image Processing,

    Z. Tu et al., “MAXIM: Multi-Axis MLP for Image Processing,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA: IEEE, June 2022, pp. 5759–5770. doi: 10.1109/CVPR52688.2022.00568

    arXiv 2022

  57. [58]

    HINet: Half Instance Normalization Network for Image Restoration,

    L. Chen, X. Lu, J. Zhang, X. Chu, and C. Chen, “HINet: Half Instance Normalization Network for Image Restoration,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Nashville, TN, USA: IEEE, June 2021, pp. 182–192. doi: 10.1109/CVPRW53098.2021.00027

    arXiv 2021

  58. [59]

    CycleISP: Real Image Restoration via Improved Data Synthesis,

    S. W. Zamir et al., “CycleISP: Real Image Restoration via Improved Data Synthesis,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA: IEEE, June 2020, pp. 2693–2702. doi: 10.1109/CVPR42600.2020.00277

    arXiv 2020

  59. [60]

    Multi-Stage Progressive Image Restoration,

    S. W. Zamir et al., “Multi-Stage Progressive Image Restoration,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA: IEEE, June 2021, pp. 14816– 14826. doi: 10.1109/CVPR46437.2021.01458. Appendix A. Github Repository This appendix will be released in a future update. We are preparing additional analyses, exte...

    arXiv 2021

  60. [2740]

    doi: 10.1109/ICIP.2019.8803391

    arXiv 2019